Axelera AI Starts Shipping Its Europa Chip, a European Challenge to Nvidia in Inference

Dutch chipmaker Axelera AI announced on September 15 that it has started shipping its second-generation inference chip, Europa, already validated on Dell and Supermicro servers. It's one of the most concrete European bets on offering a real alternative to Nvidia specifically in enterprise inference — not in training large models.

What Europa is and what formats it ships in

Europa is Axelera's second-generation inference accelerator, unveiled conceptually in October 2025 and now shipping commercially. It's available in three forms: as a bare chip for customers to design their own boards around, as a half-height, half-length PCIe card (Axelera Edge 232p), or as a full-height, full-length PCIe card (Axelera Server 250p) built for data center servers.

Technical specs and validation with manufacturers

Each Europa chip packs eight second-generation AI processing cores alongside sixteen RISC-V vector cores, delivers 629 TOPS of compute performance, and supports INT4, INT8, and INT16 precision. The Edge 232p variant is already available in validated servers from Dell (XE5) and Supermicro (111AD), and Axelera counts Advantech, Axiomtek, HPE, Lenovo, and Seco as additional hardware partners.

Why it targets inference, not training

Unlike competitors trying to replicate Nvidia's dominance in both training and inference, Axelera is deliberately positioning itself in the enterprise inference segment: running already-trained models efficiently in terms of cost and power draw, rather than competing for the market of training frontier models from scratch. It's a more accessible niche for a new player, since inference has less extreme interconnect and memory requirements than training massive models, and the supporting software is simpler to mature.

What it means if you buy or use AI

For companies in Guatemala and Latin America, Axelera Europa isn't a direct purchase option today — initial availability is centered on Europe and on integrators like Dell and Supermicro that primarily serve that market. But if your company runs high-volume inference on proprietary or open-weight models (not training), it's worth starting to track the ecosystem of accelerators specialized in inference as an alternative to general-purpose Nvidia GPUs: they tend to offer better cost per inference and lower power consumption for stable production workloads, though with less software flexibility than the CUDA ecosystem.